Fork of MineDojo/NitroGen with real-time improvements
NitroGen is an open foundation model for generalist gaming agents. This multi-game model takes pixel input and predicts gamepad actions.
- Real-time processing — real FPS without interruptions
- Simplified launch via Docker
- Improved game window search
- Use any window resolution
- Faster image processing
- Enhanced CUDA support
- Process ID support — specify process IDs instead of just names
| Component | Requirement |
|---|---|
| OS | Windows 11 |
| Python | 3.12 |
| GPU VRAM | ~10 GB (tested on RTX 5060 Ti 16GB) |
| Game | Your own copy (not distributed) |
git clone https://github.com/dffdeeq/NitroGen-real-time.git
cd NitroGen-real-time
# Download model from HuggingFace
huggingface-cli download nvidia/NitroGen --local-dir ./models ng.ptDocker handles CUDA setup automatically — no manual configuration needed.
docker compose up --build --force-recreate
# use -d flag to run in the backgroundInstall dependencies for play.py:
pip install .[play]Requires manual CUDA setup on your system.
pip install uv
# Choose your CUDA version (cu126, cu128, cu129, cu130)
uv sync --extra cu129Start the game, then run the agent:
python scripts/play.py --process '<game_executable_name>.exe'| Parameter | Description |
|---|---|
--process |
Game executable name or process ID |
--width |
Game capture width (default: 1920) |
--height |
Game capture height (default: 1080) |
--no-record |
Disable video recording (better performance) |
--no-debug-save |
Disable debug saves |
--actions-per-step |
Actions per inference (default: 18) |
For maximum performance:
python scripts/play.py --process '<game>.exe' --no-record --no-debug-save --actions-per-step 18- Open Task Manager (
Ctrl+Shift+Esc) - Right-click the game process → Properties
- Copy the name from General tab (ends with
.exe)
For processes running inside other executables (like Minecraft in javaw.exe), use PowerShell:
Get-Process <process_name> | Select-Object Id, ProcessName, MainWindowTitle | Format-Table -AutoThis project is strictly for research purposes and is not an official NVIDIA product.
